Automatic Tuberculosis Screening Using Cough

Inspiration

Tuberculosis is one of the most common causes of death in the world, and part of the reason is that there is a high patient load and a limited amount of healthcare facilities and diagnostics tools.

What it does

This project uses cough sounds as a potential biomarker for TB diagnosis. We built a model that can predict the TB status of a patient using a half-second cough recording and clinical information, including age, gender, height, and weight.

How we built it

We trained the model using LightGBM, an open-source gradient boosting framework and feature extraction. The features we chose were Mel-Frequency Cepstral Coefficients (MFCC), the delta and delta2 of MFCC, and Chroma.

Challenges we ran into

We weren't able to create an interactive website like we had wanted to due to the time limit, so we used some examples from the dataset instead. We also weren't able to train the model using deep learning, which might have increased the accuracy

Accomplishments that we're proud of

An accomplishment we're proud of is achieving an AUC of 0.732 and combining cough recordings with clinical information to predict TB.

What we learned

We learned about what feature extraction is and how it can be used in machine learning to improve the accuracy of a sound-based model.

What's next for Automatic Tuberculosis Screening Using Cough

In the future we plan to expand the data set and incorporate more features, to make the model more accurate. We plan to implement a feature to the website where users can upload their own cough recordings and clinical data and the model can predict their TB status. Another possible future direction is expanding the model to more diseases.

Built With

• python
• scikit-learn